TOPIC Universal Outlier Hypothesis Testing  
AREA Communications & Networks  
SPEAKER Dr Sirin Nitinawarat from Qualcomm Technologies Inc, San Diego 
DATE 19 December 2014, Friday 
TIME 11:00 am to 12 noon 
VENUE Communications & Networks Laboratory, E4-06-12 Department of Electrical & Computer Engineering, NUS  
FEES No Charge 

In universal outlier hypothesis testing, we seek to efficiently discover a few ``outlier'''' streams from a large number of data streams, most of which are ``typically'''' distributed. Apart from being distinct, the outlier and typical distributions can be arbitrarily close. In addition, no training data are available to the tester to learn the outlier or typical distribution. Applications of outlier hypothesis testing include event detection and environment monitoring in sensor networks, understanding visual searches by humans and animals, and fraud detection and anomaly detection in massive data.
We propose a test for both the fixed sample size and sequential settings, with the flavor of the generalized likelihood test and the repeated significance test, respectively. We show that when an outlier is present, the test is {\em universally exponentially consistent,} yielding an exponential decay in the error probability. In addition, we derive an easily computable lower bound on the achievable error exponent that applies to any fixed number of streams. The bound shows that for certain cases, our test is {\em universally asymptotically efficient} in that it achieves a limiting error exponent, as the number of streams grows unboundedly, that is equal to the best possible exponent when the outlier and typical distributions are both known.

We also describe an experiment that provides a numerical evaluation for the performance of the proposed test on real data set. The experiment could be relevant to spam detection applications.

We conclude with some relevant current and future research problems.

This is joint work with Yun Li and Prof. Venugopal V. Veeravalli at the University of Illinois.


Sirin Nitinawarat obtained the B.S.E.E. degree from Chulalongkorn University, Bangkok, Thailand, with first class honors, and the M.S.E.E. degree from the University of Wisconsin, Madison. He received his Ph.D. degree from the Department of Electrical and Computer Engineering and the Institute for Systems Research at the University of Maryland, College Park, in December 2010. Between 2011 and 2014, he was a postdoctoral research associate in the Coordinated Science Laboratory at the University of Illinois at Urbana-Champaign. He is now a senior engineer at Qualcomm Technologies, Inc., in San Diego. His research interests are in statistical signal processing, estimation and detection, information and coding theory, communications, stochastic control, and machine learning. 


 | Back to seminar list |

All are welcome to attend these Seminars. Should you have any enquiries, please contact Ms. Lily Png at 6516 6509 or you can e-mail to Ms Lily Png ().